Google announced on October 13, 2026 that Gemini 4 Argon was opening to developers, businesses and paying users after a restricted access phase. The model is available through the paid API, and the company has begun rolling it out in the Gemini app and AI Mode in Search for Google AI Ultra and Pro subscribers. This is an expansion of actual access, rather than another unveiling of an already announced model.
For anyone waiting to try Argon at work, the main change is the opportunity to choose a suitable entry point. But broad availability does not mean everyone gets it free, every tool comes with the same subscription, or every account already shows the option. The announcement divides access by product and customer type, alongside a time-limited introductory price.
What has changed since restricted access
In the original announcement, published in late September and covered by AI24 in early October, Google initially opened the model mainly to trusted cyber defenders through Fairwind. Broader routes now join that program: developers can build with the paid Gemini API in AI Studio and Android Studio, while Google AI Ultra subscribers can also try it in Google Antigravity and AI Studio.
App users face a different arrangement. Google names the Ultra and Pro plans for Gemini and AI Mode in Search. That does not promise that a Pro account also receives every benefit mentioned for Ultra in development tools. Anyone planning a particular use should match their intended workspace to the plan's terms, rather than relying on the model name alone.
Businesses also have several entry points, including Gemini for business, the API and Antigravity. This distinction matters because chatting in an app and integrating a model into a product serve different needs. The former provides a ready-made interface; the latter also requires system integration, usage management and behavior testing. The launch expands the options without removing the implementation work.
Google describes a global rollout, but its announcement does not separately confirm every combination of country, language, plan and feature. It therefore does not establish that every capability is available in Israel or has been tested in Hebrew. For an Israeli team, the practical checks are whether the feature appears in the relevant account and whether it handles the team's language tasks well.
The lower price has an expiry date
The public model's introductory price is 2 dollars per million input tokens and 10 dollars per million output tokens. Google specifies a 95% discount on input tokens read from cache. These are API usage rates, not the price of an app subscription. Confusing those payment types can produce a faulty estimate even when the individual numbers are correct.
The introductory period ends on November 30, 2026. Prices of 4 dollars per million input tokens and 20 dollars per million output tokens are scheduled to follow. A budget for a short October trial is therefore not enough to estimate ongoing December costs. The same usage at the regular input and output rates will then cost 2 times as much.
As a purely arithmetic example, a million uncached input tokens plus a million output tokens would cost 12 dollars during the launch period and 24 dollars afterward. This is not an estimate for an average task: different tasks read different amounts of material and generate different lengths of output. The example shows why both directions should be measured, rather than counting requests alone.
The cache discount is not a blanket reduction on the bill either. It applies to eligible input retrieved from cache, rather than everything sent to the model or the output it generates. It may be especially relevant to teams reusing similar background documents, but actual savings depend on the share of eligible usage. Check the selected service's billing and storage rules before budgeting.
More room for sustained work, without a guaranteed result
Google describes a unified limit of a million input and output tokens through the API and directs developers to a new continuation protocol. Unified matters: the figure should not be read as a promise of a million incoming tokens plus another million answer tokens in every request. It is a working envelope that integrations need to understand and plan around.
Google positions Argon for complex coding, scientific and multistep professional tasks. For a development team, the interesting question is not just how much text the model can process, but whether it preserves project requirements through revisions, tests and intermediate failures. A larger working window can support a longer sequence; it does not certify that the last change is correct.
One way to test this is a bounded task with a verifiable outcome: modifying a code component with existing tests, or summarizing documents against a defined question list. A useful comparison keeps the input material and success criteria constant. This separates an impressive-looking answer from work that saves time after a person has checked, corrected and approved it.
Google is also inviting scientists to apply for the highest Antigravity usage tier free for 6 months. The program is intended for 10,000 scientists, with rolling onboarding subject to eligibility, capacity and selection. It is an opportunity to apply, rather than an automatic entitlement for every researcher or a promise of immediate access after submitting the form.
Public access is separate from the special cyber program
Alongside the wider launch, Google is extending Fairwind to hundreds of additional partners. That program is for trusted cyber defenders and provides access without the public version's cyber guardrails. The distinction matters: paying for the public API does not make a customer a program member or automatically grant the same level of access.
In its announcement, Google describes training to respect permission boundaries, seek approval before irreversible steps and monitor agent actions. These are safeguards presented by the company, not a guarantee that a connected system will never make a mistake. An organization allowing a model to work with files, services or accounts still needs to decide which actions are permitted and which require approval.
A useful starting point is therefore a small project with limited permissions, a defined budget and recorded results. Measure not only how long the model worked, but also the time needed to check it and the corrections required. If it produces more work while increasing review demands, the business benefit may be smaller than the initial result on screen suggests.
The novelty of this launch is clear: Argon moves from a model most users could only read about to one they can begin incorporating into work through the announced routes. Adoption now depends on 3 separate questions: suitable access, quality on the real task and cost after the discount ends. Test all of them in the same trial before expanding usage.